PDCA54: compare one actual parser behavior request

한국어

One new request bounds JSONL event-envelope keys to64 entries and128 decoded UTF-8 bytes, rejecting duplicates. Sol6/Luna low each ran twice. All four candidates passed75 independent terminal tests, exited zero and reported complete core usage. The tests and repetitions do not add unique requests.

What was fixed before outcomes

Plan54 was committed at 063f274f2a3a1776b793447cdb4b51be09a93743 before outcomes: Sol6/Luna/Luna/Sol6, concurrency1, main120s plus independent verification45s, without executor retry/resume/fallback. Only the existing local ChatGPT login was used; no new API key, paid job or endpoint was created.

The executor was built cleanly from CI-passed 4373f9f6b20dba922b1cf5309d079e50a4591aeb; the public task base and contract bind revision ee72334166e2962b0821d5198a50fdd13f92ab29. The new external-repository contracts were not used for this fixed comparison.

The native Laya prediction was also committed before coding, at 5aae164b788c638ad6b9f296b7346567f68cb977. Unchanged INT8 base, CPU1thread and threshold0.9 proposed standard with score 0.610351, therefore abstained. Configured fallback was strong/Astra; Astra was not executed. One cold whole-process observation took 1,183ms, app observation1,151.896ms, maximum RSS 1,508,032,512bytes (about1.51GB/1.40GiB). This includes native loading; it is not warm latency, Go heap or GPU resources. The score is not calibrated coding-success probability, and this memory does not meet the ultra-small-resource target.

Actual observations

Order	Requested profile	Main ms	Verification ms	Input	Included cached input	Output	Included reasoning	Outcome
1	Sol6 low	66,431	4,750	163,569	141,568	2,036	48	Accepted, zero exit, complete usage
2	Luna low	28,981	4,470	78,701	66,304	663	0	Accepted, zero exit, complete usage
3	Luna low	28,488	4,503	85,762	71,424	665	0	Accepted, zero exit, complete usage
4	Sol6 low	61,574	4,730	157,577	144,256	1,918	109	Accepted, zero exit, complete usage

Input485,609 includes cached423,552; output5,282 includes reasoning157. Sum of main times185,474ms and separate verification18,453ms excludes parent preparation. Luna had smaller reported input/output and observed main time in both repetitions of this request. Uncached input sums were Luna26,735 versus Sol6 35,322. Subscription consumption, billing and internal provider calls/retries remain unknown, so these numbers do not establish monetary savings.

Numeric records and hashes match each private executor terminal record.json byte-for-byte. All four observed process-group cleanup and removal of copied auth; candidate-written tests did not determine acceptance. Requested models are explicit, while provider-attested served identity remains unknown. Sources 1/4 are identical; sources2/3 differ from each other. Passing a shared contract does not mean identical implementations.

Integration into the real tool

The first zero-exit accepted mutable pkg/taskoutcome/summary.go was integrated unchanged, SHA256 64bbcfd17d3321b74f1b679142baa79906ded759458a245ade6a546172e7deb5. A fixed [64]string array resets per event and checks decoded-key count, UTF-8 byte length and duplicates. Nested items remain opaque, and unknown names/values are not retained in Summary. No shared state or locks were added.

The six previously frozen independent external-package tests were copied into runtime regression with test names changed only. Candidate-authored tests were not adopted. Independent review confirmed the final missing-type condition is equivalent under the original type validation. Error precedence when one input violates both a size and duplicate rule is outside the frozen assessment. This code change does not validate routing utility or cost savings.

The first Linux CI exposed an incorrect test expectation: a formatter-normalized UUID baseline without the new API requires isolated compilation, so a platform without supported isolation must report verifier_unknown rather than rejection. The test now checks that unknown outcome and actual isolated rejection on macOS separately. Product decisions, frozen specs, parser observations and original source files remain unchanged.

The next macOS CI reached the default ten-minute timeout for the test package that aggregates multiple cold isolated compiler checks. Its current control had run for four seconds. Only the outer CI race-suite budget is explicitly twenty minutes, with a thirty-minute test-job cap. Each candidate still has the same 45-second default / 60-second maximum verification timeout, ten-second inner test timeout, output limits, isolation and acceptance criteria. This change is not a model execution-time or performance observation.

Scale and next steps

Actual cumulative development observations now cover 5 distinct requests, 16 CLI records,3 code families and1 repository. Earlier failures, support errors and timeout remain unchanged. The8 local candidates and 120 public Go candidates are neither execution nor label counts. Independent contracts are prepared for2 of those120 external candidates; external model attempts, training and final eligibility remain0.

24 development probes cannot justify general performance claims. Expand to120+ independently verifiable development requests across repositories and240+ requests for usable training-signal assessment. Separately acquire and protect at least 2,400 distinct final requests per claimed domain. Previously read requests, repetitions, translations and sibling tasks do not inflate final counts. The scale guide and acquisition plan describe grouping and minimum execution budgets. This cycle performs no fitting, calibration, new weight release, production activation or protected-final scoring.